RobustBench-TC - Misleading Description: leaderboard
RobustBench-TC (2026) tests tool-calling models under deployment noise grounded in real tool-calling failures, using 199 single-turn samples drawn from BFCL, API-Bank, RoTBench, ToolAlpaca and ToolEyes and scored by deterministic rule-based parsers. This board is tool-call accuracy in percent under a misleading description on the correct tool and an attractive budget or fast suffix on a distractor; higher is better.
Metric: Tool-call accuracy (%) under the MisDesc perturbation (reward component: misleading description on GT; _Budget/_Fast suffix on distractor); RobustBench-TC samples drawn from BFCL V3 single-turn, API-Bank, RoTBench, ToolAlpaca and ToolEyes, each scored by its source benchmark's strict scorer after a format-tolerant tool-call parse, temperature 0 (DeepSeek-R1-Distill at 0.6), Qwen3-family models with thinking disabled; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | O4 Mini | 60.8 | #368 | |
| 2 | DeepSeek R1 Distill Qwen 14B | 42.2 | #2771 | |
| 3 | Qwen 3.5 9B | Non-reasoning | 41.2 | #828 |
| 4 | MUA-RL-14B | 38.2 | #1658 | |
| 5 | LoopTool-8B | 37.3 | #1473 | |
| 6 | MUA-RL-32B | 35.3 | ||
| 7 | Qwen 3 32B | Non-reasoning | 34.3 | #972 |
| 8 | Qwen 3 14B | Non-reasoning | 33.3 | #1269 |
| 9 | LoopTool-32B | 33.3 | ||
| 10 | Qwen 3 8B | Non-reasoning | 32.4 | #1873 |
| 11 | MUA-RL-8B | 32.4 | #1640 | |
| 12 | ToolRL-Qwen2.5-1.5B | 26.5 | #2299 | |
| 13 | Qwen 2.5 1.5B Instruct | 22.5 | #7470 | |
| 14 | Qwen 2.5 3B Instruct | 20.6 | #5392 | |
| 15 | TL-CodeLLaMA-2 | 19.6 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=robustbench-tc-misleading-description · How It Works · Data refreshed daily, snapshot 2026-11-21.